Kimi K3 on a Mac mini? Show me the bytes.

Video caption

Kimi K3 on a Mac mini? Show me the bytes. Disk offload and sparse experts do not create bandwidth. Compact local for bounded work; cluster for full K3. #EricFieldNotes

Full written post / accessibility read

Start with Kimi's two point eight trillion parameters. At four bits per parameter, raw weights alone are about one point four terabytes. The maximum current Mac mini has sixty-four gigabytes of unified memory. That is before scales, cache and runtime overhead.

Someone may show the model files on a drive or a single generated token. K3's sparse design reduces how much computes per token, but the full expert weights still need storage and movement. The user cares about first-token delay, output rate and concurrency, not a successful launch banner.

Specify the exact hardware and quantized build. Run representative prompts under cold and warm conditions; record first-token delay, steady tokens per second, memory pressure, concurrent throughput, electricity and accepted-task quality. Compare with a hosted K3 route and a smaller local model.

Use a compact local model when it meets the task contract. Use a fabric-scale service when you need full K3 capability at useful throughput. Do this because parameter access is not a substitute for memory bandwidth, serving topology and measured quality.

#EricFieldNotes

Four-beat scene transcript

1. Kimi K3 on a Mac mini? Show me the bytes.

Start with Kimi's two point eight trillion parameters. At four bits per parameter, raw weights alone are about one point four terabytes. The maximum current Mac mini has sixty-four gigabytes of unified memory. That is before scales, cache and runtime overhead.

Visual: The full weight floor is already over a terabyte.

2. A download proves almost nothing.

Someone may show the model files on a drive or a single generated token. K3's sparse design reduces how much computes per token, but the full expert weights still need storage and movement. The user cares about first-token delay, output rate and concurrency, not a successful launch banner.

Visual: Disk offload and sparse experts do not create bandwidth.

3. Measure a real service envelope.

Specify the exact hardware and quantized build. Run representative prompts under cold and warm conditions; record first-token delay, steady tokens per second, memory pressure, concurrent throughput, electricity and accepted-task quality. Compare with a hosted K3 route and a smaller local model.

Visual: Warm and cold delay, sustained rate, concurrent users.

4. Buy the right capacity.

Use a compact local model when it meets the task contract. Use a fabric-scale service when you need full K3 capability at useful throughput. Do this because parameter access is not a substitute for memory bandwidth, serving topology and measured quality.

Visual: Compact local for bounded work; cluster for full K3.

Research and claim limits

• Moonshot AI: Kimi K3 official model card (S128)
• Apple: current Mac mini technical specifications (S129)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.